feat(linkedin-studio): N8 — stemme-kjeden (chronicle-voice-profile + drift-logg + language-resolve + auto-gull) [skip-docs]

Bind the long-form VOICE before prose — the discipline the line already applies to SOURCES (validate-binding), its strongest link. - C-1/A2-F1: voice-trainer distils a chronicle-voice-profile from the published gold corpus (per language/format); newsletter Step 0/3a/3b bind drafting to it, with an explicit never-silent fallback to the short-form corpus (which the engine itself calls degrading for chronicle voice). - C-2: Step 3a Pass 0 READS the chronicle-voice-drift-log as a negative list (it was append-only, never consumed); Step 3b re-reads it in expansion. - C-3: Step 1 resolves + persists the top-level edition-state.language, making the template doc-string's "Resolved at Step 1" claim true; the Norwegian review axis (language-reviewer / voice-scrubber) now fires when the edition IS Norwegian instead of being silently off by default. Step 6.5 already consumed the field — resolving it closes the axis end-to-end. - A2-F2: Step 8 lock appends the locked edition to ingest/published (provenance=published), so the gold corpus the profile distils from grows one edition at a time — never the engine's drafts. Domain-general throughout (languages/formats from the user's own corpus). test-runner 140 -> 146 (Section 16o: 6 unconditional voice-chain greps + non-vacuity self-test; anti-erosion floor 123 -> 129). All six suites green (trends 300, brain 134, hooks 140, tests 35, render 60). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014bE7VbkmR3cqHFEeGfzgwb
2026-07-23 22:58:38 +02:00 · 2026-07-23 22:58:38 +02:00 · b30b0e208f
commit b30b0e208f
parent 9de38c4939
4 changed files with 197 additions and 13 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@ -16,6 +16,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **`arcs` CLI verb + `/linkedin:trends --demand` mode** — render the §4 arc map (`node --import tsx src/cli.ts arcs`), reachable via the command's new demand mode (delegates to `demand-spotter`, keeps commands at 30). The morning brief also shows the per-candidate demand signal alongside the existing reader fields.
 - Trends suite 276 → 300; test-runner 139 → 140; `tsc --noEmit` clean.

+### Added — voice chain: bind voice BEFORE prose (N8 / C-1, C-2, C-3, A2-F1, A2-F2)
+
+- **`chronicle-voice-profile` artifact (C-1/A2-F1)** — `voice-trainer` now distils a dedicated long-form voice profile (`${DATA}/profile/chronicle-voice-profile.md`) from the **published gold corpus only** (`ingest/published/`, `provenance=published`), one section per `(language, format)` pair present in the corpus — never from the English short-form `authentic-voice-samples.md` the engine itself calls degrading for chronicle voice, never from `provenance=ai-draft`. `/linkedin:newsletter` binds long-form drafting (Step 0 / 3a / 3b) to THAT profile as the primary reference, with an **explicit, never-silent fallback** to the short-form corpus when it is absent.
+- **Drift log is READ, not just written (C-2)** — Step 3a gains a **Pass 0**: read `chronicle-voice-drift-log.md` as a *negative list* of the recurring AI-tells/drifts `voice-scrubber` logged on prior editions, so drafting avoids them BEFORE the scrub (the log was append-only, never consumed); Step 3b re-reads it during expansion.
+- **Review language resolved + persisted at Step 1 (C-3)** — the calibration now **resolves the review `language`** (detect/ask) and persists the top-level `edition-state.language` at the Step 1.5 checkpoint, making the template's "Resolved at Step 1" doc-string true. Previously the default `"en"` was never set, so `language-reviewer`'s Norwegian checks and `voice-scrubber`'s gold standard silently never fired even for a Norwegian edition; Step 6.5 already consumed the field, so resolving it closes the axis end-to-end.
+- **Auto-gull — the gold corpus grows on its own (A2-F2)** — Step 8 lock now appends the locked edition to `ingest/published/` tagged `provenance=published`, so the corpus the chronicle profile is distilled from grows one edition at a time (the author's own locked editions, never the engine's drafts). No manual curation.
+- Domain-general throughout: languages/formats come from the user's own corpus, never hardcoded. test-runner 140 → 146 (Section 16o: 6 unconditional voice-chain greps + self-test; anti-erosion floor 123 → 129).
+
 ### Built feedback (awaiting consumer-side proof)

 - **MR-F9** (demand-sweep «innenfra og ut») — built; the (a)/(b)/(c) evidence gate is a runtime demonstration, proven consumer-side (plugin agents don't resolve in the dev repo).
--- a/agents/voice-trainer.md
+++ b/agents/voice-trainer.md
@ -152,6 +152,39 @@ Architecture: [prose/sectioned/framework]

 Flag if: fewer than 5 samples (low confidence), single time period (temporal bias), or contradictory patterns (possible ghostwriting).

+### Long-form chronicle voice profile (per language / format)
+
+Short-form feed posts and long-form editions (newsletters, essays, chronicles) are
+**different registers of the same author** — a device the author bans in a punchy
+short post (e.g. the em-dash) can be load-bearing rhythm in a long-form chronicle.
+Drafting long-form against the short-form corpus therefore *degrades* the chronicle
+voice (this is `voice-scrubber`'s canonical rule). So the long-form pipeline binds
+to its own distilled artifact, which this agent generates:
+
+- **Artifact:** `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/profile/chronicle-voice-profile.md`
+  — a distilled long-form voice profile built with the same six-dimension
+  framework as the general profile, but from **only the long-form gold corpus**.
+- **Corpus (published gold only):** the approved/published long-form editions under
+  `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/ingest/published/`
+  carrying `provenance=published`. The same model-collapse fence applies —
+  **never distil chronicle voice from `provenance=ai-draft`**, and never from the
+  short-form `voice-samples/authentic-voice-samples.md` corpus (wrong register).
+- **Per language AND format — domain-general.** Emit one profile section per
+  `(language, format)` pair actually present in the corpus (e.g. `no` / newsletter),
+  keyed by those axes; the author's languages and formats come from their OWN
+  corpus, never hardcoded. If a `(language, format)` pair has no published editions
+  yet, say so plainly and emit no section for it — do not invent a chronicle voice
+  from short-form samples or from assumptions.
+- **Grows automatically.** Each locked edition is appended to the gold corpus at
+  `/linkedin:newsletter` Step 8 (`provenance=published`), so re-running this
+  distillation sharpens the chronicle profile over editions without manual curation.
+
+Trigger this distillation on the same events as the general profile (a new
+published edition, quarterly audit, explicit request). The long-form pipeline
+(`/linkedin:newsletter` Step 0 / 3a / 3b) reads this artifact; when it is missing,
+that pipeline warns and falls back — it never silently drafts against the wrong
+corpus.
+
 ## Voice Drift Detection

 ### Drift Scoring
--- a/commands/newsletter.md
+++ b/commands/newsletter.md
@ -176,9 +176,24 @@ the edition left off before doing anything.
   Step 2. Do not confuse `<serie>/STATE.md` (this edition's production state)
   with the plugin's own `STATE.md` / `docs/BUILD-HANDOVER.local.md` (which govern
   building the plugin itself).
-4. **Read the voice profile** — `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/authentic-voice-samples.md`
-   and anything else under `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`. Long-form must match the
-   author's voice; this is the reference for every drafting and review phase.
+4. **Read the long-form voice profile** — the **chronicle-voice-profile**
+   (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/profile/chronicle-voice-profile.md`),
+   the distilled chronicle voice for THIS edition's language/format, built by
+   `voice-trainer` from the published gold corpus. This is the **primary reference
+   for every long-form drafting and review phase**. Short-form feed posts and
+   long-form editions are different registers of the same author, so long-form must
+   **not** be drafted against the English short-form corpus
+   (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/authentic-voice-samples.md`) —
+   that corpus encodes rules (e.g. no em-dash) that are wrong for chronicle voice
+   and would degrade it. Read the `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`
+   corpus as a **secondary tributary** (register cues), never as the primary
+   long-form reference.
+   **Fallback (never silent).** If `chronicle-voice-profile.md` is absent — or has
+   no section for this edition's language/format — say so plainly, note the
+   chronicle voice is uncalibrated, and fall back to today's behavior (draft against
+   `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`), while
+   flagging the gap so the operator can seed the gold corpus (Step 8 auto-gull grows
+   it from each locked edition).
 5. **Resolve the active personas (per-artifact).** Personas are configured **per
   edition**, not from one fixed global file. Resolve the set for
   `articles.<currentArticle>` in this order (Step 1 finalizes + persists it):
@ -258,12 +273,12 @@ Then display a short status:
 Edition: <series title> — article <currentArticle> "<title>"
 Last completed phase: <currentPhase>   (or: NEW edition)
 Resuming at: Step <N> — <step name>   (from the resumption table above)
-Voice profile: loaded | MISSING
+Chronicle voice profile: loaded (<language>/<format>) | MISSING → fallback to short-form corpus
 Persona library: <N> personas loaded (active set chosen in Step 1)
 ```

-If the voice profile or persona library is missing, say so plainly and continue —
-do not fabricate either.
+If the chronicle voice profile or persona library is missing, say so plainly and
+continue on the Step-0 fallback — do not fabricate either.

 ## Step 1: Brief + calibration

@ -321,6 +336,16 @@ Settle these dimensions (most should come from context, not questions):
 - **Tone** — respected-peer vs. teaching-down; calibrated to the primær.
 - **Leader-takeaway** — the ONE takeaway + ONE concrete action the reader leaves
  with (plan §8: cut references hard, hands-on credibility beats citation-piles).
+- **Language** — **Resolve the review language** for this edition: detect it from
+  the series brief / prior editions / the drafting language, else ask once. Persist
+  it to `edition-state.json` → `language` (e.g. `"no"` for Norwegian, `"en"` for
+  English) at the **Step 1.5 checkpoint**. This is not cosmetic — it gates the
+  language-dependent review agents: `language-reviewer`'s Norwegian-idiom /
+  «kanselli-stil» checks and `voice-scrubber`'s gold standard fire against THIS
+  language, and never grade prose against Norwegian idiom when the edition is not
+  Norwegian. Left unresolved, the template default `"en"` silently disables the
+  Norwegian axis even for a Norwegian edition (the C-3 bug this closes) — so resolve
+  it explicitly, do not assume the default.

 Use `AskUserQuestion` only for the genuinely open dimensions (cap ≈3). Good
 candidates: which personas are in scope + which is primær; the angle if the
@ -446,6 +471,11 @@ Edition brief
   - If Step 1 ran the **trend-intake bridge**, persist `articles.NN.sourceTrendId`
     (the trend id) here too — the durable provenance link Step 10 reads to auto-`act`
     the trend. Omit it (leave `null`) for a manually-started edition.
+   - Persist the resolved **`language`** (Step 1) — write it to the **top-level**
+     `language` (the series/edition-level review language; there is no per-article
+     `language` field) so the language-dependent review agents grade against the
+     right language from here on. This is the write that makes the template
+     doc-string's "Resolved at Step 1" claim true.
   - Write `articles.NN.livedSpecifics` (the slot-map above, `status: "pending"`).
   - Set `currentPhase: "lived-specifics"` in `edition-state.json`.
   - Write a "lived-specifics bound → next: research (scoped by the binding)" line
@ -771,9 +801,21 @@ Typically ~20–30 % of the edition's final length.

 **Procedure:**

-1. **Re-read the voice profile** (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`) before writing a
-   single sentence — this is the existing LTL rule and it is not optional for
-   long-form. Voice match starts at the spine, not at expansion.
+1. **Pass 0 — load the voice reference + the drift negatives** before writing a
+   single sentence (voice match starts at the spine, not at expansion — the
+   existing LTL rule, not optional for long-form):
+   - Re-read the **chronicle-voice-profile**
+     (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/profile/chronicle-voice-profile.md`)
+     for this edition's language/format — the reference the spine must match
+     (fall back to `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`
+     per Step 0 if it is missing).
+   - Read the **chronicle-voice-drift-log**
+     (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/chronicle-voice-drift-log.md`)
+     as a **negative list**: the recurring AI-tells and voice drifts `voice-scrubber`
+     logged on prior editions. Do NOT reproduce them in the spine — the log is no
+     longer write-only; it steers drafting BEFORE the scrub, so the Step-4/persona
+     passes and the scrubber find less each edition. If the log is absent, note it
+     and proceed.

 2. **For each section in `NN-skjelett.md`, write ONE paragraph that delivers
   that section's pitch.** No examples yet, no anecdotes, no research citations
@ -865,8 +907,11 @@ turning-points the spine already named.

 **Procedure:**

-1. **Re-read the voice profile** (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`) before expanding —
-   the voice was set at the spine; do not lose it in expansion.
+1. **Re-read the chronicle-voice-profile**
+   (`${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/profile/chronicle-voice-profile.md`,
+   fallback `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/voice-samples/`)
+   and the drift-log negatives (Step 3a Pass 0) before expanding — the voice was set
+   at the spine; do not lose it, and do not let expansion reintroduce a logged drift.

 2. **Expand section by section, against the spine.** Each section's paragraph
   from Step 3a is the *contract* for that section: the expansion must
@ -1707,6 +1752,19 @@ produces the editor's single delivery artifact — `POST.html`, the
   body is frozen — Step 9 may only adjust the distribution hook, never the
   locked edition text.

+5. **Auto-gull — grow the gold corpus (A2-F2).** A locked edition is now
+   published-grade voice signal, so append it to the gold corpus the chronicle
+   voice profile is distilled from:
+   - Record the locked body (`NN-utkast.md`) as a new record under
+     `${LINKEDIN_STUDIO_DATA:-$HOME/.claude/linkedin-studio}/ingest/published/`,
+     tagged **`provenance=published`** (never `ai-draft`) with this edition's
+     `language` and format in its frontmatter, so `voice-trainer`'s
+     chronicle-voice-profile distillation (per language/format) picks it up.
+   - This closes the loop the model-collapse fence protects: the corpus the
+     chronicle voice is learned from is the author's OWN locked editions, growing
+     one edition at a time — never the engine's drafts. The next edition's Step 0
+     therefore binds to a sharper profile without any manual corpus curation.
+
 ```
 LOCK → delivery.
 - Preconditions: factcheck 🔴 = none, persona resonans primær = JA   (else: STOP)
--- a/scripts/test-runner.sh
+++ b/scripts/test-runner.sh
@ -1512,6 +1512,88 @@ fi

 echo ""

+# --- Section 16o: Voice Chain — chronicle-voice-profile binding (N8 / C-1,C-2,C-3,A2-F1,A2-F2) ---
+echo "--- Voice Chain (N8) ---"
+
+# N8 binds the long-form VOICE before prose — the same discipline the line already applies to
+# SOURCES (validate-binding), its strongest link. Five invariants must hold in the .md/.json surfaces:
+#  (C-1/A2-F1) voice-trainer.md distils a chronicle-voice-profile from the PUBLISHED gold corpus, and
+#              newsletter.md binds drafting (Step 0/3a/3b) to THAT profile — not to the English short-form
+#              authentic-voice-samples corpus the engine itself calls degrading for chronicle voice.
+#  (C-2)       Step 3a READS the chronicle-voice-drift-log as a Pass-0 negative list (today write-only).
+#  (C-3)       Step 1 RESOLVES + persists the review `language`, making the edition-state doc-string's
+#              "Resolved at Step 1" claim true, so the Norwegian axis fires when content IS Norwegian.
+#  (A2-F2)     Step 8 lock appends the locked edition to the gold corpus (ingest/published,
+#              provenance=published) so the corpus the profile distils from grows automatically.
+# All are UNCONDITIONAL .md/.json greps (no TS deps → they raise the deps-absent floor). Auto-gull needs
+# BOTH gold-corpus literals; a non-vacuity self-test guards that compound predicate (mirrors Section 16c).
+CVP_LIT='chronicle-voice-profile'
+DRIFTLOG_LIT='chronicle-voice-drift-log'
+LANG_RESOLVE_LIT='Resolve the review language'
+GOLD_PATH_LIT='ingest/published'
+GOLD_PROV_LIT='provenance=published'
+
+autogull_wired() {  # $1 = text; wired iff BOTH gold-corpus literals present (echo twice — grep consumes stdin)
+  echo "$1" | grep -qF "$GOLD_PATH_LIT" && echo "$1" | grep -qF "$GOLD_PROV_LIT"
+}
+
+VC_SELFTEST_OK=1
+if ! autogull_wired "lock appends the edition to ingest/published with provenance=published"; then
+  VC_SELFTEST_OK=0; echo "  non-vacuity FAIL: a fully-wired auto-gull probe was not detected"
+fi
+while IFS= read -r probe; do
+  [ -z "$probe" ] && continue
+  if autogull_wired "$probe"; then
+    VC_SELFTEST_OK=0; echo "  false-positive FAIL: under-wired auto-gull probe accepted -> $probe"
+  fi
+done <<'NEGATIVE16O'
+lock appends the edition to ingest/published/ but never tags provenance
+marks provenance=published but never names the gold-corpus path
+NEGATIVE16O
+if [ "$VC_SELFTEST_OK" -eq 1 ]; then
+  pass "voice-chain self-test: auto-gull predicate needs BOTH gold-corpus literals (1 accepted, 2 under-wired rejected)"
+else
+  fail "voice-chain self-test failed — the auto-gull lint is vacuous or over-eager"
+fi
+
+NL="commands/newsletter.md"
+VT_N8="agents/voice-trainer.md"
+
+# (C-1/A2-F1) chronicle-voice-profile artifact generated + bound
+if grep -qF "$CVP_LIT" "$VT_N8"; then
+  pass "voice-trainer.md distils the $CVP_LIT (per-language gold-corpus voice artifact)"
+else
+  fail "voice-trainer.md missing '$CVP_LIT' — the chronicle voice artifact is never generated (C-1/A2-F1)"
+fi
+if grep -qF "$CVP_LIT" "$NL"; then
+  pass "newsletter.md binds long-form drafting to the $CVP_LIT (C-1)"
+else
+  fail "newsletter.md missing '$CVP_LIT' — drafting still binds only to the short-form corpus (C-1)"
+fi
+
+# (C-2) drift-log READ (Pass 0 negative list), not just written
+if grep -qF "$DRIFTLOG_LIT" "$NL"; then
+  pass "newsletter.md reads the $DRIFTLOG_LIT as a Pass-0 negative list (C-2)"
+else
+  fail "newsletter.md missing '$DRIFTLOG_LIT' — the drift log is still write-only (C-2)"
+fi
+
+# (C-3) language resolved + persisted at Step 1
+if grep -qF "$LANG_RESOLVE_LIT" "$NL"; then
+  pass "newsletter.md Step 1 resolves + persists the review language ('$LANG_RESOLVE_LIT', C-3)"
+else
+  fail "newsletter.md missing the Step 1 language-resolution instruction ('$LANG_RESOLVE_LIT', C-3)"
+fi
+
+# (A2-F2) auto-gull at Step 8 lock — the gold corpus grows on its own
+if autogull_wired "$(cat "$NL")"; then
+  pass "newsletter.md Step 8 auto-gull: locked edition appended to the gold corpus ('$GOLD_PATH_LIT' + '$GOLD_PROV_LIT', A2-F2)"
+else
+  fail "newsletter.md missing auto-gull — lock does not grow the gold corpus (needs '$GOLD_PATH_LIT' + '$GOLD_PROV_LIT', A2-F2)"
+fi
+
+echo ""
+
 # --- Section 18: Assertion-Count Anti-Erosion (SC6) ---
 # The lint self-modifies its own checks, so a green run could mask a silently dropped
 # assertion. Pin the total pass()+fail() invocations as a monotonic floor; the count
@ -1540,12 +1622,15 @@ echo ""
 # cli.ts saturation-at flag grep) = 117; +6 for RE-R3e's six UNCONDITIONAL Section-16n checks
 # (brief-diff self-test + brief.ts diffSurfaced grep + brief.ts parseSurfacedFrontmatter grep +
 # brief.ts 'Nytt siden sist' section grep + cli.ts selectPriorBriefFile wiring grep + brief.ts
-# 'surfaced: ' frontmatter grep) = 123.
+# 'surfaced: ' frontmatter grep) = 123; +6 for N8's six UNCONDITIONAL Section-16o checks (voice-chain
+# self-test + voice-trainer chronicle-voice-profile grep + newsletter chronicle-voice-profile grep +
+# newsletter chronicle-voice-drift-log read grep + newsletter Step-1 language-resolve grep + newsletter
+# Step-8 auto-gull compound grep) = 129.
 # NB: the floor tracks the deps-absent MINIMUM (conditional TS suites warn-skip and drop
 # the count), so it is bumped only by UNCONDITIONAL new checks — NOT pinned to the
 # deps-present TOTAL_CHECKS (that would zero the warn-skip margin and false-fail a fresh
 # clone). Runs last so TOTAL_CHECKS sees every prior check.
-ASSERT_BASELINE_FLOOR=123
+ASSERT_BASELINE_FLOOR=129
 TOTAL_CHECKS=$((PASS + FAIL))
 if [ "$TOTAL_CHECKS" -ge "$ASSERT_BASELINE_FLOOR" ]; then
  pass "assertion-count anti-erosion: $TOTAL_CHECKS checks >= baseline floor $ASSERT_BASELINE_FLOOR"